Four Factors (Dean Oliver): shooting eFG%, ball control TOV%, rebounding ORB%, and FTr (free-throw rate) — the left block is a team's offense, the right block (shaded) is what they allow on defense. This is *why* a team's rating is what it is.

WAR — wins above replacement, v4: our blend of box scores, play-type efficiency, playmaking, rim protection and hustle defense. Offensive and defensive components are standardized and combined, then converted to season value using minutes and replacement level. This is a descriptive player-value estimate, not a probability of winning MVP or a verified betting edge. Off / Def — the offensive and defensive components in points/100 above average. Shot+ — points over expected on the same shots. Floor / Ceil — 20th/80th-percentile scoring nights (props reliability). Clutch — points in the last 5 min within 5, with each player's shooting lift vs. their season eFG%.

How it's built: a gradient-boosted minutes model (recent minutes + rest + who's out — a teammate's absence redistributes minutes and usage) feeds per-stat models that add recent form and a shot-volume vs. shooting-efficiency split. The strongest single lever we found is teammate availability: minutes MAE dropped from 4.87 to 4.60 once the model sees who's sitting. Opponent-specific signals — team defense, pace, even exact matchup — were tested and add almost nothing.

Def Rating — points saved per matchup vs. what the offensive players "should" score, standardized across the league (higher = better). We residualize out minutes and assignment difficulty, because raw matchup numbers reward bench players hidden on weak scorers. Honest caveat: this is on-ball defense only — it doesn't capture rim protection, help, or scheme, so it's a descriptive lens, not a full defensive rating.

Reading it: Off Freq = share of the offense's possessions in that play type; Off PPP = their points per possession on it; Def PPP = what the defense allows on it; vs Lg = how the defense compares to league average (green = stingy); Matchup = the offense's efficiency shifted by this defense's strength on that play type. An honest note we found in backtests: this grid is a superb scouting tool, but at the team level it doesn't out-predict the simple rating difference — great defenses tend to be great everywhere.

Reading it: bigger cell = more shots from that spot; amber = the team scores more points-per-shot there than the league average from that range, blue = less. A big amber cluster behind the arc is an elite, high-volume shooting team.

Impact = the team's average point margin per game with the player minus without (positive = better with them). Click a player to see how teammates' scoring changes when they sit. Caveat: players miss few games, so big numbers on a tiny "G out" are noisy — this is descriptive, not a controlled on/off (which needs lineup data).

How it works: win probability = 1 / (1 + 10^(−ΔElo/400)), where ΔElo includes the home-court bonus. Projected margin = ΔElo / 28. This is our Stage-1 baseline; the bottom-up matchup simulator (next) will project individual player and team stat lines.

What makes it a projection (next): each line gets adjusted for the specific opponent — pace, defensive rating faced, and the deep per-player splits (perimeter defense, shooting vs. taller/shorter defenders) — then regressed with shrinkage so small samples don't mislead.

NBAI reads from a three-layer parquet pipeline (raw → clean facts → leakage-safe features) — the same data engine behind every rating, prediction, and metric on this site.